Kafunk VS pulsar-client-dotnet

Performance.

Generally speaking, F# was actually very fast, and had nice concurrency support, but there were times that wasn't the case.

For example, in 2016 I was part of the initiative to rewrite the ad feed. We had to read in several Kafka topics, do some joining on our end, and emit to a separate Kafka topic. This isn't terribly hard to write, but we were dealing on the order of about ~100gb of data being pushed into memory. This is hardly "big data" stuff, but it's enough to highlight some issues.

Specifically, the built F# persistent map structure was simply too slow to get the performance we wanted. I really like that structure, it's really handy and nice, but I ended up having to make heavy use of the ConcurrentDictionary that was built into .NET. This wasn't that hard or anything, but it made me a little sad that I had to move to a mutable store to get the performance I needed.

There was also the fact that the `async` monad, while generally very good and useful, had bizarre bottlenecks that were hard to measure. It was difficult to know when the async task was actually started, and when you tried to measure performance bottlenecks you were really only measuring the scheduler, not the actual performance. This isn't really F#'s fault, this is an issue with any kind of cooperative scheduling system, but occasionally to get the performance we needed we'd have to move to lower level threads instead of the pretty monadic stuff. Microsoft eventually released the Task monad which generally performed a bit better.

There were other things here and there; the Kafka client libraries for .NET simply aren't as good as the Java ones. Jet actually open-sourced their own (https://github.com/jet/kafunk) which did make it a bit more functional and nice, but it had performance issues as well, so a lot of us ended up using Confluent.

There were little annoyances specific to F# as well; there's no real concept of a monad transformer, so if you wanted to do something like, for example, combine an Option and an Async into generalized syntax, you'd have to write your own wrapper monad thing, which wasn't that hard but was sort of ad hoc.

The general rule of thumb was that the first draft of software, we would try and keep as functional and pretty. If that was too slow, we allowed mutation but only within a function. If that was too slow, we'd allow global mutation but only with thread-safe stuff.

Isn't not having perfect support with Apache is, in a way, a criticism of any non-JVM language? Perhaps the problem is on the other end?

We need a couple more Apache Struts and Log4J incidents for the public to realize there may be better technologies and ecosystems. Spring is very slow anyways.

As for Pulsar, there is a nice and optimized client (written in F#!) https://github.com/fsprojects/pulsar-client-dotnet